fix: restore throughput timeline hydration and Gemini timeouts

[RobertTLange]

What changed

• centralize LLM timeout and retry policy across providers, and convert Gemini client timeouts to the millisecond unit expected by google-genai
• restore async runner slot reservation and completed-job persistence safety, with added regression coverage for duplicate persistence and slot reuse edge cases
• stop persisting sampling/evaluation worker lane ids and fall back to timestamp-based throughput lane inference for runtime views
• disable browser caching for the local WebUI HTML shells and fix Throughput tab hydration so right_tab=throughput restores reliably after data loads
• improve the Throughput runtime timeline UI with adaptive per-worker row sizing and a default-visible Hide Plot / Show Plot toggle
• improve the local WebUI UX by adding dashboard sorting controls and surfacing the active results directory in the meta header
• fix the compare view so negative best scores remain visible instead of being clamped away by summary rendering
• update the unreleased changelog to cover the latest runtime and WebUI changes

Why

• Gemini requests were timing out too aggressively because the Google client interpreted the shared timeout value in milliseconds, not seconds
• async runs were vulnerable to slot-accounting and duplicate-persistence regressions under retries and overlapping completions
• the throughput runtime timeline was both fragile and hard to read: hydration could fail after load, worker lanes could regress, and dense runs compressed multiple workers into unreadable rows
• the local WebUI still had a few workflow rough edges around stale cached JS, result discovery, dashboard scanning, and compare summaries

How to test

• pytest -q tests/test_llm_client_backends.py tests/test_llm_retry_constants.py tests/test_gemini_provider.py
• pytest -q tests/test_async_runner_recovery.py tests/test_pipeline_timing.py tests/test_runtime_timeline_webui.py
• pytest -q tests/test_compare_webui.py tests/test_index_webui.py
• uv run ruff check tests --exclude tests/file.py
• uv run mypy --follow-imports=skip --ignore-missing-imports tests/test_*.py tests/conftest.py
• uv run --with pytest-cov pytest -q -m "not requires_secrets" --cov=shinka --cov-report=term-missing --cov-report=xml:coverage.xml

Context

• the branch now combines backend safety fixes, timeout normalization, and a set of local WebUI quality-of-life fixes needed to make throughput analysis and local monitoring reliable again
• commits are kept split by concern so reviewers can inspect runtime safety, WebUI behavior, and changelog/docs updates independently